Category: 4. Technology

Xreal, the Chinese augmented reality start-up behind Google’s first Android XR-powered augmented reality (AR) glasses, is gearing up for the mass-market release of its lightweight eyewear as early as the first quarter of 2026, its co-founder said.

Code-named Project Aura, the product was first announced at Google’s annual I/O developer conference in May. Rivalling offerings from Apple and Meta Platforms, Google adopted a strategy from its smartphone playbook: providing the operating system – Android XR – while collaborating with hardware makers.

Project Aura is Google’s second XR initiative, following Moohan, a collaboration with Samsung Electronics on a virtual reality headset similar to Apple’s Vision Pro.

Aura’s field of view (FoV) – the visible area seen through the glasses – exceeds 70 degrees, “the largest screen we have ever made”, according to a statement from Xreal at the Augmented World Expo in June. The device will be powered by a Qualcomm Snapdragon chipset alongside Xreal’s own XR-optimised X1S chip.

“This achievement is thanks to Xreal’s extensive experience in optical engineering, enabling thinner lenses with broader fields of view,” Wu Kejian, the company’s co-founder, chief scientist and head of algorithms, said in an interview with the South China Morning Post.

Recent research indicates that LLMs, particularly smaller ones, frequently struggle with robust reasoning. They tend to perform well on familiar questions but falter when those same problems are slightly altered, such as changing names or numbers, or adding irrelevant but related information. This weakness, known as poor out-of-distribution (OOD) generalization, results in notable accuracy drops, even in simple math tasks. One promising solution is to create synthetic variations of reasoning problems, helping models learn to focus on the underlying logic rather than surface details. Strengthening reasoning in this manner is crucial for developing more general and reliable AI systems.

Abstracting the Core Logic of LLM Reasoning Failures

LLMs have demonstrated impressive reasoning capabilities, yet they often falter when exposed to distribution shifts, such as changes in phrasing, numerical values, or the introduction of distractions. This vulnerability is evident across benchmarks in logic, mathematics, and commonsense reasoning. Prior solutions have relied on data augmentation to expose models to a broader variety of inputs, improving robustness but increasing computational demands. Researchers have also explored formats such as abstraction-of-thought and chain-of-abstraction to teach abstract reasoning, while planning techniques like chain-of-thought and tree-of-thought aid step-by-step problem-solving. Reinforcement learning and preference-based methods provide additional support for reasoning skill development beyond pattern memorization.

AbstRaL’s Symbolic Learning Method to Improve Reasoning Consistency

Researchers from Apple and EPFL propose AbstRaL, a method that teaches LLMs to understand abstract reasoning patterns rather than memorizing surface details. Instead of generating many varied training examples, which is computationally costly, AbstRaL helps LLMs learn the underlying structure of reasoning problems using reinforcement learning. This method connects these abstract patterns to symbolic tools, enabling more reliable problem-solving. Tested on GSM benchmarks, AbstRaL significantly improves LLM performance, especially when faced with input changes or distracting information. It outperforms models trained only with supervised learning by promoting more consistent and context-independent reasoning.

Four Steps to Abstract Symbolic Reasoning via AbstRaL

AbstRaL is a four-step framework designed to teach LLMs to reason abstractly rather than rely on surface patterns. First, it identifies key variables in a question and replaces them with symbolic placeholders. Then, using specially crafted data (GranulAR), the model learns to reason step-by-step with these abstract symbols. Next, it retrieves the general reasoning structure (abstraction) from the symbolic answer. Finally, it uses this abstraction with the original values to compute the correct answer. Reinforcement learning with two rewards, one for correctness and another for symbolic similarity, further improves the model’s ability to generate accurate, context-independent reasoning patterns.

GSM8K Variations Reveal AbstRaL’s Robustness Across LLM Sizes

The researchers evaluate AbstRaL on math reasoning tasks using models such as Llama-3 and Qwen2, training them with a dataset called GranulAR that rewrites math problems in an abstract symbolic form. This helps models focus on structure rather than surface details. They test robustness using altered versions of GSM8K problems, changing numbers, names, and phrasing. Compared to baselines like standard Chain-of-Thought prompting, AbstRaL shows stronger consistency and less accuracy drop on these variations. Especially for smaller models, it improves reliability across reworded inputs. The results suggest that teaching models to reason abstractly makes them more adaptable and less reliant on memorized patterns.

Teaching LLMs Abstract Thinking through Reinforcement Yields Robust Reasoning

In conclusion, AbstRaL is a method designed to enhance abstract reasoning in LLMs, making them more resilient to superficial changes in problems. Unlike traditional fine-tuning or data augmentation, AbstRaL uses reinforcement learning to train models on GranulAR rationales that mix Socratic chain-of-thought with detailed abstraction. This approach helps models strip away surface-level distractions and better connect with symbolic tools. Tested on challenging GSM8K perturbation benchmarks, AbstRaL notably reduces performance drops under distribution shifts, particularly in smaller models. The study shows that learning to abstract improves reasoning robustness more effectively than relying solely on direct supervision.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter, Youtube and Spotify and don’t forget to join our 100k+ ML SubReddit and Subscribe to our Newsletter.

Sana Hassan, a consulting intern at Marktechpost and dual-degree student at IIT Madras, is passionate about applying technology and AI to address real-world challenges. With a keen interest in solving practical problems, he brings a fresh perspective to the intersection of AI and real-life solutions.

Sent out today by longtime Freedreno/MSM open-source Qualcomm GPU driver developer Robin Clark are the main set of MSM kernel graphics/display driver updates targeting the upcoming Linux 6.17 merge window. There are several exciting feature additions coming to this next kernel version for those relying on Qualcomm graphics capabilities.

First up, the MSM kernel graphics driver is landing VM_BIND support. VM_BIND can help with lowering CPU overhead for the user-space Mesa drivers to deliver better performance. Additionally, VM_BIND is needed for Vulkan sparse requirements. The VM_BIND support for the Qualcomm MSM driver has been in development for a while and amounts to dozens of patches with the main focus on being supporting the Vulkan sparse memory needs by the TURNIP Mesa driver.

Another notable feature addition is providing Qualcomm Adreno X1-85 GPU “speedbin” support. Speedbin is important for reaching the maximum performance capabilities on Snapdragon graphics processors. The Adreno X1-85 is interesting for being the graphics found on the Qualcomm Snapdragon X Elite laptops. Long story short, with Linux 6.17 the graphics performance will hopefully be nicely improved thanks to the speedbin functionality now in place.

This open-source driver code also has Snapdragon X1-45 GPU support for that graphics processor found within the Snapdragon X Plus SoCs.

The MSM driver changes for Linux 6.17 also now provide a single source of truth for Universal Bandwidth Compression (UBWC) configuration handling, further decoupling between the GPU and KMS code, SM8750 support in the DPU and DSI and MDSS code, and various fixes.

More details on these pending Qualcomm MSM DRM driver improvements slated for Linux 6.17 via today’s pull request to DRM-Next. I’ll be working on some new Qualcomm Snapdragon X Elite benchmarks with the Acer X1E laptop once the Linux 6.17 cycle is underway.

I’ve been upgrading my home network lately, and while I was figuring out how some of the wired connections would go, I opted for RJ45 SFP+ modules because some of my devices have 10GbE NICs. I could have gone with a mix of Ethernet and fiber, but I wanted everything to be the same, so it’d be easier to install and have some spare parts.

I was wrong. I should have gone for the mix, because I noticed some issues with the Ethernet runs. More specifically, the ones that use RJ45 SFP+ modules, where the transceiver gets way too warm for comfort even when idle. But I’ve also noticed issues with the 10GbE NICs on my motherboards glitching. If that’s through heat, it’s not fixable; if it’s driver issues, I’m not equipped to do that either, so it’s time to add some new adapters.

I’ve never been so glad to have ATX motherboards with extra PCIe slots. I’ve added ex-enterprise networking cards and am now using SFP+ active optical DACs for the main connections around my office, and I couldn’t be happier. I’ll add more to the mix wherever I can, but I’m also looking at fiber runs with modules to have more flexibility over speeds and upgrades.

Ethernet is fine, really

But RJ45 transceivers get super hot and I hate it

I’m not knocking Ethernet here, as with the correct CAT specification, I could get faster speeds than I already have. It’s not about speed, at least not here; it’s about reliability and temperature. Those are intertwined anyway, as the rising temperatures of my 10GbE ports and modules are affecting the stability and reliability of my network.

How much of an effect is harder to quantify. I could set up network monitoring and see how many packets have to be re-sent, or do some deep performance analysis and get hard numbers, but I don’t have the time or inclination for my home lab. Not when I can fix the issue for a few bucks, which is what I did. If I had more than a few devices to upgrade, then I’d have run the numbers because the upgrade cost would have been more substantial.

Some of my network gear has RJ45 10GbE but not enough ports

I do have some hardware on my network with 10GbE-capable RJ45 ports, but they’re few and far betweenand are currently wired to my NAS and a Wi-Fi 7 access point. Some of my computers have 2.5 GbE ports and need a new NIC anyway, so why not get fiber-capable ones? The only piece I haven’t added yet is an all-SPF+ switch, but that’ll be the next thing that goes in, and then most of my office will be fiber runs.

It was time for a change

I now have faster speeds, and future upgrades are cheaper

While looking for replacement network cards, I prioritized reliability over all else. If some level of upgradeability was available, that would be nice, too. Going to 10GbE was already a big jump, but I didn’t want to have to upgrade again if I hit the network’s limits. I wasn’t sure if I would, as the only thing hitting anywhere near 10GbE was my RAID when dealing with big file transfers, but after not planning the last network upgrade properly, I needed to do this the right way.

10GbE SPF+ cards with one or two ports are fairly inexpensive, but they mostly use Intel chipsets, and I noticed many forum threads discussing issues. Not what I wanted to hear (or deal with!), but non-Intel cards were expensive. Or so I thought, as I headed to eBay, I found hundreds of older Mellanox adapters that were capable of 10/25GbE, with two ports, and for less than the Intel-based cards.

A couple of purchases (and days) later, I had new-to-me networking cards and active optical DACs to replace the copper wires I was using. I knew I might run into firmware issues with the cards, but I had a list of resources to flash them with the correct firmware, and it turned out I didn’t need them as they got recognized straight away.

Now I’ve got a more stable network, with less heat in my office, and optical cables that are much easier to route than the thick Ethernet cables I was using. Plus, it taught me a bit more about networking, including how enterprise hardware is easier to work with in terms of firmware flexibility, which was a nice surprise. And with every card having two ports, I can use them to link between my desktops, making it less of a hassle to run the cables.

Related

5 reasons I’m still hard-wiring Ethernet even when I have a Wi-Fi 7 AP

Wi-Fi is pretty fly but I’m a consistency guy.

Even though I love these optical DACs for my office, I still have a few Ethernet runs I can’t remove

I’ve now got the bulk of my bandwidth-hogging devices on SFP+ active optical DACs, but I can’t switch out a few cable runs because I need PoE++ to power things like access points and other switches. Eventually, I’ll pull fiber through and replace the Cat5e that’s in my walls, but that day is far off. I need time to plan where power is coming from and save up for the eventual electrician bill. But until then, I know that the devices that need fast, stable networking are all handled by lasers, and I love the thought of that.

Samsung’s Galaxy Z Flip 7 looks to be a bigger surprise than initially expected, with the first hands-on leak confirming reports of a bigger cover display and thinner hardware, and it looks good.

A hands-on video of the Galaxy Z Flip 7 was posted on Twitter/X this weekend, showing off the device in plenty of detail. We can see the device, in black, with its side profile looking a fair bit thinner compared to prior generations. The open device is also seen with virtually no crease, suggesting this is probably a brand new device that just came out of the box, with the packaging directly behind it suggesting that’s the case.

That video, which was since been removed, also shows the cover display in action for the first time.

There’s a new widget panel that takes advantage of the larger display with four visible widgets at once. The screenshots captured by SamMobile, unfortunately, don’t show much more than that. Still, it’s great to see the screen lit up.

Advertisement – scroll for more content

We were able to obtain a copy of another video (thanks Yuval!) which only shows the hardware. Rumor has it the Flip 7 is a fair bit thinner than Samsung’s past few generations, a change that we’ve been waiting quite a while to see.

Samsung has reservations open now for the Galaxy Z Fold 7 and Flip 7, with a $50 credit when you pre-order and up to $1,150+ in savings. Reservations are free, and there’s no obligation to actually buy a device if you sign up, but there’s no other time you can get that $50 credit (which can go towards Galaxy Watch 8), 3x reward points, and other perks.

More on Samsung:

Follow Ben: Twitter/X, Threads, Bluesky, and Instagram

Digital Wellbeing tools aren’t designed only to help you manage your screen time and sleep habits, but they are also there to raise your awareness. Sometimes, you might pick up your phone, unlock it, and start fiddling with an app without realizing the app you’re using. I do this all the time with Reddit, where I subconsciously open it.

According to a 2025 survey conducted by Harmony Healthcare IT (1,000 respondents), Americans spend an average of 5 hours and 16 minutes per day on their phones. That’s around a 14% increase compared to data gathered in 2024. That equates to more than half a working or school day spent on our phones.

Unfortunately, digital addiction is a real thing. That’s why Google brought Digital Wellbeing tools to help combat this ongoing issue. Thankfully, you can freely enable these tools through your Android settings, whether you own a Pixel device or another budget option; you can also leverage these tools to gain some surprising insights into your personal phone habits.

6

Learn how many notifications you get in a day

Find out which apps are giving you the most grief

One reason we feel almost obligated to check our phones is due to the notifications. Unless you explicitly set up your notifications to distinguish between essential and less important ones, you’re likely to check your phone to find out where it is from. Notifications can be a primary reason why you feel the need to be on your phone, as many apps demand your attention.

When you want to gain insight into your alerts, Google’s Digital Wellbeing can provide that for you. Google’s Digital Wellbeing tools tell you how many notifications you receive per day. It even tells you which apps are responsible for sending them, as you might find that some apps are more disruptive or demanding of your attention than others, and uninstalling or muting them may be the key to reducing your screen time.

5

Learn what apps you use most daily

Some apps are far more addictive than others

The most obvious insight you gain from Google’s Digital Wellbeing tools is learning which apps you use the most. The neat part is that you can understand this through a graph (sorted by day or hourly). Sometimes, it might not even be about a particular third-party app. You’d be surprised about your activity in general. For me, I spend a considerable amount of time adjusting my device’s settings and browsing the Google Play Store, which I wouldn’t have known if I hadn’t looked into the data.

It also tells you how long you’re accessing them. To me, it’s essential to know what’s dictating my screen time since I’d rather limit the lengthy, unimportant apps than the ones I check daily but only spend a few minutes with.

4

Learn what websites you visit daily

Web browsing consumes more time than you think

Aside from learning about the apps you use, you can learn about the websites you visit daily if you’re using Chrome. When you visit the activity page, you can add permissions to show your Chrome activity in Digital Wellbeing.

When you check your activity, the websites you’ve visited are displayed. It’s also split by how long you spend browsing those websites. Similar to app activity, your activity can be sorted by the amount of time spent daily or hourly. I found it interesting that I duplicated my activity between the website version and some apps, which helped me determine which content type captured my attention the most.

3

Learn when you are getting the most screen time

You can discover some interesting patterns

Aside from learning websites and apps that take up the most screen time. You can also learn more when you’re most active. For example, I spend most of my screen time in the evenings, and I am particularly active on weekends. That’s partly because it’s when I take breaks from work (around dinner time and late in the evening) and also when I read the news and social media. I’ve also learned that I spend a considerable amount of time on my phone during my lunch break, as it’s far easier to browse on a smaller device while I’m eating.

Gaining insight into my screen time habits, particularly when I use my phone excessively, helps me become more aware of my behavior. Even now, since I spend a lot of unnecessary time browsing on it over lunch, it motivates me to try and change that by picking up a paperback book to read instead.

2

Learn how many times you open an app

You might be opening apps you don’t even use

Not only do Google’s Digital Wellbeing tools tell you about the apps and websites you use, but you can also learn more about how often you open up an app. When you select the View activity details option, you can choose Times opened from the dropdown menu above the graph. Above the graph, it will tell you the total number of unlocks you complete in a day or week. Below the graph, you will find the apps that you’ve opened. Choosing an app from the list will produce a separate graph that can be broken down by the time you’ve opened that particular app, either hourly or daily.

I find this insight particularly interesting, given how many unlocks I perform on my phone. For example, I wasn’t the biggest believer in enabling Always on Display on my phone. I thought it was just contributing to battery drain and wasn’t useful enough to keep it on. However, using it has helped me reduce the time I spend using my phone.

My most important notifications appear below the time and date on my Always on Display. As a result, leaving the Always on Display on makes it so that I am far less tempted to unlock my phone to check my apps. So, Google’s Digital Wellbeing would reflect that I unlock my phone less and open fewer apps than before due to my change in habits.

1

Learn about your sleeping

Bedtime mode helps with sleep health awareness

Source: Unsplash

Bedtime mode is a built-in Digital Wellbeing tool that helps you improve your sleep habits by limiting your screen time. You won’t gain statistics that tell you about your sleeping habits — you’d want to pair it with a sleep-tracking app or use an actual sleep tracker for that. But how you use Bedtime mode is still telling.

For example, when I started using Bedtime mode to reduce my screen time and lower the number of notifications I receive, I found that I fell asleep 30-40 minutes earlier than before. Sometimes, I slip and browse Reddit right before bed to read TV and movie discussions, and it has kept me up longer than I intended. To me, Bedtime mode made a noticeable difference in my sleeping habits.

If you find that you’re still picking up your phone during Bedtime mode, you can also set up screen time reminders to encourage you to stay off your apps while checking your phone. Bedtime mode is excellent at helping you gain awareness of your sleep and the actions you need to take to improve it.

Related

How to set up and use Bedtime mode on your Android phone

Use Bedtime mode effectively to reduce distractions and get better sleep

Your eyes could use the break

Balancing screen time between your devices can be challenging, especially if your job requires it. But your eyes can use the break. When you spend a lot of time looking at screens, it strains your eyes. But if you can’t physically separate from your phone or computer because your livelihood depends on it, consider adding blue light filters and following the 20-20-20 rule.

Running DOOM on ARM? Old hat. Running Doom (2016) on an ARM SBC? Well, that’s a bit more interesting, and [Interfacing Linux] shows us how with this handy guide, and in a video embedded below.

The ARM SBC in question is the Radxa Orion O6, which claims to be the world’s first open source ARMv9 motherboard. It has a 12-core ARM9.2 SoC from Cix, and crucially for gaming, a PCIe x16 slot to fit a graphics card. There’s onboard graphics in the SoC, of course, but that’s only going to cut it for old DOOM, not Doom (2016).

Obviously there is no ARMv9 build of this nine-year-old classic. Thus, a certain degree of hacking is required. An x86 emulation layer called FEX is used to run Steam, which includes the Proton compatibility layer to sit between the Linux system and the Windows system calls the games are making.

So, again: Windows games, running via a compatibility layer on a Linux system running on an x86 emulator, running on another Linux system, running on ARM hardware. Yeah, we’d call that a hack, and just possibly cursed. The amazing thing is that it works. As long as you use a PCI 3.0 graphics card, anyway; apparently newer ones don’t work, or at least not the one [Interfacing Linux] had to try. (Some report better luck with newer NVIDIA cards.) Which is a pity, because every game tested is GPU-throttled on this system. Considering the CPU should be handling all the emulation, that just goes to show how GPU-heavy even 10-year-old games are.

[Interfacing Linux] seems to enjoy running things where they should not; we just wrote up their SteamOS console.

This isn’t the first time a Radxa board has shown up on Hackaday either, but they don’t seem to be nearly as popular as the fruit-flavoured competition for the sort of projects we usually see. If you’ve got a good project using one of these powerful ARM boards– or anything else, don’t hesitate to send us a tip!